Creator / Lead Engineer · 2025 – present
Prismatic Engine
Open-source multi-agent orchestration engine (AGPLv3) that runs a production fleet of AI agents operating two live businesses daily — with mechanical verification on every output.
The problem
AI agents can do real operational work — triage, research, code changes, business follow-up — but only if a human can trust the output. Unchecked agent work fails silently: tasks stall, claims go unverified, and the operator ends up babysitting the system that was supposed to help.
The build
- Architected agent decomposition, orchestration, and review systems end to end: work is broken into owned tasks with acceptance criteria, executed by specialized agents, and reviewed before it lands.
- Built human-in-the-loop control into the design: every agent output passes mechanical verification — tests, counts, evidence — before it counts as done.
- Ran a measured fleet-reliability program: baselined 213 production agent runs, then built the fixes (a budget-checkpoint plugin, automated review sweeps, playbook routing) against that baseline.
- Packaged the operating patterns as open tooling other operators can run: the pstack agent playbooks and the hermes-muse-bridge, both public under MIT.
The evidence
- The fleet runs two live businesses daily — it is production infrastructure, not a demo.
- Fleet baseline measured across 213 production runs: 46% completion, 65% first-attempt success — the numbers the reliability program is improving against.
- Budget-checkpoint plugin shipped with a 49-test suite.
- kanban-stream, a work-streaming plugin written for this stack, was merged upstream into NousResearch's hermes-agent (October 2026).
- Companion repos public on GitHub: hermes-muse-bridge (v1.5.0) and muse-pstack (v1.1.0), MIT licensed.
Prismatic Engine is the system I built to make AI agents dependable enough to run real operations. It decomposes work into tasks with explicit acceptance criteria, routes them to specialized agents, enforces budget checkpoints so long tasks report before they run out of runway, and puts a verification pass between an agent’s claim and anything being marked done.
The design principle throughout: trust comes from evidence, not from the agent’s summary. Test counts, file diffs, and measured baselines are the review inputs — the same discipline the engine’s own reliability program runs on.